1/*
2 *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include <immintrin.h>
12
13#include "./vpx_dsp_rtcd.h"
14#include "vpx/vpx_integer.h"
15#include "vpx_dsp/x86/bitdepth_conversion_avx2.h"
16#include "vpx_ports/mem.h"
17
18static void hadamard_col8x2_avx2(__m256i *in, int iter) {
19  __m256i a0 = in[0];
20  __m256i a1 = in[1];
21  __m256i a2 = in[2];
22  __m256i a3 = in[3];
23  __m256i a4 = in[4];
24  __m256i a5 = in[5];
25  __m256i a6 = in[6];
26  __m256i a7 = in[7];
27
28  __m256i b0 = _mm256_add_epi16(a0, a1);
29  __m256i b1 = _mm256_sub_epi16(a0, a1);
30  __m256i b2 = _mm256_add_epi16(a2, a3);
31  __m256i b3 = _mm256_sub_epi16(a2, a3);
32  __m256i b4 = _mm256_add_epi16(a4, a5);
33  __m256i b5 = _mm256_sub_epi16(a4, a5);
34  __m256i b6 = _mm256_add_epi16(a6, a7);
35  __m256i b7 = _mm256_sub_epi16(a6, a7);
36
37  a0 = _mm256_add_epi16(b0, b2);
38  a1 = _mm256_add_epi16(b1, b3);
39  a2 = _mm256_sub_epi16(b0, b2);
40  a3 = _mm256_sub_epi16(b1, b3);
41  a4 = _mm256_add_epi16(b4, b6);
42  a5 = _mm256_add_epi16(b5, b7);
43  a6 = _mm256_sub_epi16(b4, b6);
44  a7 = _mm256_sub_epi16(b5, b7);
45
46  if (iter == 0) {
47    b0 = _mm256_add_epi16(a0, a4);
48    b7 = _mm256_add_epi16(a1, a5);
49    b3 = _mm256_add_epi16(a2, a6);
50    b4 = _mm256_add_epi16(a3, a7);
51    b2 = _mm256_sub_epi16(a0, a4);
52    b6 = _mm256_sub_epi16(a1, a5);
53    b1 = _mm256_sub_epi16(a2, a6);
54    b5 = _mm256_sub_epi16(a3, a7);
55
56    a0 = _mm256_unpacklo_epi16(b0, b1);
57    a1 = _mm256_unpacklo_epi16(b2, b3);
58    a2 = _mm256_unpackhi_epi16(b0, b1);
59    a3 = _mm256_unpackhi_epi16(b2, b3);
60    a4 = _mm256_unpacklo_epi16(b4, b5);
61    a5 = _mm256_unpacklo_epi16(b6, b7);
62    a6 = _mm256_unpackhi_epi16(b4, b5);
63    a7 = _mm256_unpackhi_epi16(b6, b7);
64
65    b0 = _mm256_unpacklo_epi32(a0, a1);
66    b1 = _mm256_unpacklo_epi32(a4, a5);
67    b2 = _mm256_unpackhi_epi32(a0, a1);
68    b3 = _mm256_unpackhi_epi32(a4, a5);
69    b4 = _mm256_unpacklo_epi32(a2, a3);
70    b5 = _mm256_unpacklo_epi32(a6, a7);
71    b6 = _mm256_unpackhi_epi32(a2, a3);
72    b7 = _mm256_unpackhi_epi32(a6, a7);
73
74    in[0] = _mm256_unpacklo_epi64(b0, b1);
75    in[1] = _mm256_unpackhi_epi64(b0, b1);
76    in[2] = _mm256_unpacklo_epi64(b2, b3);
77    in[3] = _mm256_unpackhi_epi64(b2, b3);
78    in[4] = _mm256_unpacklo_epi64(b4, b5);
79    in[5] = _mm256_unpackhi_epi64(b4, b5);
80    in[6] = _mm256_unpacklo_epi64(b6, b7);
81    in[7] = _mm256_unpackhi_epi64(b6, b7);
82  } else {
83    in[0] = _mm256_add_epi16(a0, a4);
84    in[7] = _mm256_add_epi16(a1, a5);
85    in[3] = _mm256_add_epi16(a2, a6);
86    in[4] = _mm256_add_epi16(a3, a7);
87    in[2] = _mm256_sub_epi16(a0, a4);
88    in[6] = _mm256_sub_epi16(a1, a5);
89    in[1] = _mm256_sub_epi16(a2, a6);
90    in[5] = _mm256_sub_epi16(a3, a7);
91  }
92}
93
94static void hadamard_8x8x2_avx2(int16_t const *src_diff, ptrdiff_t src_stride,
95                                int16_t *coeff) {
96  __m256i src[8];
97  src[0] = _mm256_loadu_si256((const __m256i *)src_diff);
98  src[1] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
99  src[2] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
100  src[3] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
101  src[4] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
102  src[5] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
103  src[6] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
104  src[7] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
105
106  hadamard_col8x2_avx2(src, 0);
107  hadamard_col8x2_avx2(src, 1);
108
109  _mm256_storeu_si256((__m256i *)coeff,
110                      _mm256_permute2x128_si256(src[0], src[1], 0x20));
111  coeff += 16;
112  _mm256_storeu_si256((__m256i *)coeff,
113                      _mm256_permute2x128_si256(src[2], src[3], 0x20));
114  coeff += 16;
115  _mm256_storeu_si256((__m256i *)coeff,
116                      _mm256_permute2x128_si256(src[4], src[5], 0x20));
117  coeff += 16;
118  _mm256_storeu_si256((__m256i *)coeff,
119                      _mm256_permute2x128_si256(src[6], src[7], 0x20));
120  coeff += 16;
121  _mm256_storeu_si256((__m256i *)coeff,
122                      _mm256_permute2x128_si256(src[0], src[1], 0x31));
123  coeff += 16;
124  _mm256_storeu_si256((__m256i *)coeff,
125                      _mm256_permute2x128_si256(src[2], src[3], 0x31));
126  coeff += 16;
127  _mm256_storeu_si256((__m256i *)coeff,
128                      _mm256_permute2x128_si256(src[4], src[5], 0x31));
129  coeff += 16;
130  _mm256_storeu_si256((__m256i *)coeff,
131                      _mm256_permute2x128_si256(src[6], src[7], 0x31));
132}
133
134void vpx_hadamard_16x16_avx2(int16_t const *src_diff, ptrdiff_t src_stride,
135                             tran_low_t *coeff) {
136  int idx;
137#if CONFIG_VP9_HIGHBITDEPTH
138  DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
139  int16_t *t_coeff = temp_coeff;
140#else
141  int16_t *t_coeff = coeff;
142#endif
143
144  for (idx = 0; idx < 2; ++idx) {
145    int16_t const *src_ptr = src_diff + idx * 8 * src_stride;
146    hadamard_8x8x2_avx2(src_ptr, src_stride, t_coeff + (idx * 64 * 2));
147  }
148
149  for (idx = 0; idx < 64; idx += 16) {
150    const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
151    const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
152    const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
153    const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
154
155    __m256i b0 = _mm256_add_epi16(coeff0, coeff1);
156    __m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
157    __m256i b2 = _mm256_add_epi16(coeff2, coeff3);
158    __m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
159
160    b0 = _mm256_srai_epi16(b0, 1);
161    b1 = _mm256_srai_epi16(b1, 1);
162    b2 = _mm256_srai_epi16(b2, 1);
163    b3 = _mm256_srai_epi16(b3, 1);
164
165    store_tran_low(_mm256_add_epi16(b0, b2), coeff);
166    store_tran_low(_mm256_add_epi16(b1, b3), coeff + 64);
167    store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 128);
168    store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 192);
169
170    coeff += 16;
171    t_coeff += 16;
172  }
173}
174
175int vpx_satd_avx2(const tran_low_t *coeff, int length) {
176  const __m256i one = _mm256_set1_epi16(1);
177  __m256i accum = _mm256_setzero_si256();
178  int i;
179
180  for (i = 0; i < length; i += 16) {
181    const __m256i src_line = load_tran_low(coeff);
182    const __m256i abs = _mm256_abs_epi16(src_line);
183    const __m256i sum = _mm256_madd_epi16(abs, one);
184    accum = _mm256_add_epi32(accum, sum);
185    coeff += 16;
186  }
187
188  {  // 32 bit horizontal add
189    const __m256i a = _mm256_srli_si256(accum, 8);
190    const __m256i b = _mm256_add_epi32(accum, a);
191    const __m256i c = _mm256_srli_epi64(b, 32);
192    const __m256i d = _mm256_add_epi32(b, c);
193    const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
194                                            _mm256_extractf128_si256(d, 1));
195    return _mm_cvtsi128_si32(accum_128);
196  }
197}
198